Gated Mechanism Enhanced Multi-Task Learning for Dialog Routing

Proceedings of the 29th International Conference on Computational Linguistics, pages 3064–3073 October 12–17, 2022. 3064 Gated Mechanism Enhanced Multi-Task Learning for Dialog Routing Ziming Huang1∗, Zhuoxuan Jiang1∗, Ke Wang2∗, Juntao Li3, Shanshan Feng4, Xian-Ling Mao5 1Tencent, 2Alibaba DAMO Academy, 3Soochow University, 4Harbin Institute of Technology(Shenzhen), 5Beijing Institute of Technology hzmyouxiang@gmail.com, jzhx@pku.edu.cn, wk258730@alibaba-inc.com, ljt@suda.edu.cn, victor_fengss@foxmail.com, maoxl@bit.edu.cn Abstract Currently, human-bot symbiosis dialog sys- tems, e.g., pre- and after-sales in E-commerce, are ubiquitous, and the dialog routing compo- nent is essential to improve the overall effi- ciency, reduce human resource cost, and en- hance user experience. Although most existing methods can fulfil this requirement, they can only model single-source dialog data and can- not effectively capture the underlying knowl- edge of relations among data and subtasks. In this paper, we investigate this important prob- lem by thoroughly mining both the data-to- task and task-to-task knowledge among vari- ous kinds of dialog data. To achieve the above targets, we propose a Gated Mechanism en- hanced Multi-task Model (G3M), specifically including a novel dialog encoder and two tai- lored gated mechanism modules. The proposed method can play the role of hierarchical infor- mation filtering and is non-invasive to exist- ing dialog systems. Based on two datasets col- lected from real world applications, extensive experimental results demonstrate the effective- ness of our method, which achieves the state-of- the-art performance by improving 8.7%/11.8% on RMSE metric and 2.2%/4.4% on F1 metric. 1 Introduction Traditionally, a lot of human resource cost is spent on supporting the calling/online centers for customer care, such as pre-and after-sales for E- commerce and banking. With the rapid develop- ment of AI techniques for dialog systems, various bot agents have been deployed in those scenarios to reduce parts of human workload. Both the bot and human agents constitute a new and practical symbiosis dialog system which can keep a bal- ance between service quality and human resource cost (Oraby et al., 2017). In such a human-bot symbiosis environment, it is commonly seen that users’ needs continuously ∗This work was done when they were in IBM Research China, and Zhuoxuan Jiang is the corresponding author. Dialog Router Bot Agent skilled at simple queries Human Agent A skilled at ‘Technical Issue’ category Human Agent B skilled at ‘Completion Issue’ category ① ② … user Human-bot Hybrid Dialog System User Utterance NPS Dialog Category … … … Utterance i 6.5 Finding Content Utterance i+1 5.8 Finding Content Utterance i+2 4.3 Completion Issue … … … Sorry for not being very helpful. You can’t find the course in your learning queue and the progress are all gone, correct? You did not answer my question, I can’t find the course and I want to resume it Response Module Figure 1: An example of dialog routing in human-bot symbiosis dialog system. change, or sometimes they are dissatisfactory with the current servicing agents. They would request the dialog system to replace the agents with more qualified ones, because usually each bot or human agent is trained for specific domains. For …

Content truncated to 500 words